2 * G.723.1 compatible decoder
3 * Copyright (c) 2006 Benjamin Larsson
4 * Copyright (c) 2010 Mohamed Naufal Basheer
6 * This file is part of Libav.
8 * Libav is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
13 * Libav is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with Libav; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25 * G.723.1 compatible decoder
28 #define BITSTREAM_READER_LE
29 #include "libavutil/channel_layout.h"
30 #include "libavutil/mem.h"
31 #include "libavutil/opt.h"
34 #include "acelp_vectors.h"
35 #include "celp_filters.h"
36 #include "g723_1_data.h"
39 #define CNG_RANDOM_SEED 12345
45 ACTIVE_FRAME, ///< Active speech
46 SID_FRAME, ///< Silence Insertion Descriptor frame
56 * G723.1 unpacked data subframe
59 int ad_cb_lag; ///< adaptive codebook lag
69 * Pitch postfilter parameters
72 int index; ///< postfilter backward/forward lag
73 int16_t opt_gain; ///< optimal gain
74 int16_t sc_gain; ///< scaling gain
77 typedef struct g723_1_context {
80 G723_1_Subframe subframe[4];
81 enum FrameType cur_frame_type;
82 enum FrameType past_frame_type;
84 uint8_t lsp_index[LSP_BANDS];
88 int16_t prev_lsp[LPC_ORDER];
89 int16_t sid_lsp[LPC_ORDER];
90 int16_t prev_excitation[PITCH_MAX];
91 int16_t excitation[PITCH_MAX + FRAME_LEN + 4];
92 int16_t synth_mem[LPC_ORDER];
93 int16_t fir_mem[LPC_ORDER];
94 int iir_mem[LPC_ORDER];
106 int16_t audio[FRAME_LEN + LPC_ORDER + PITCH_MAX + 4];
109 static av_cold int g723_1_decode_init(AVCodecContext *avctx)
111 G723_1_Context *p = avctx->priv_data;
113 avctx->channel_layout = AV_CH_LAYOUT_MONO;
114 avctx->sample_fmt = AV_SAMPLE_FMT_S16;
116 avctx->sample_rate = 8000;
117 p->pf_gain = 1 << 12;
119 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
120 memcpy(p->sid_lsp, dc_lsp, LPC_ORDER * sizeof(*p->sid_lsp));
122 p->cng_random_seed = CNG_RANDOM_SEED;
123 p->past_frame_type = SID_FRAME;
129 * Unpack the frame into parameters.
131 * @param p the context
132 * @param buf pointer to the input buffer
133 * @param buf_size size of the input buffer
135 static int unpack_bitstream(G723_1_Context *p, const uint8_t *buf,
140 int temp, info_bits, i;
142 init_get_bits(&gb, buf, buf_size * 8);
144 /* Extract frame type and rate info */
145 info_bits = get_bits(&gb, 2);
147 if (info_bits == 3) {
148 p->cur_frame_type = UNTRANSMITTED_FRAME;
152 /* Extract 24 bit lsp indices, 8 bit for each band */
153 p->lsp_index[2] = get_bits(&gb, 8);
154 p->lsp_index[1] = get_bits(&gb, 8);
155 p->lsp_index[0] = get_bits(&gb, 8);
157 if (info_bits == 2) {
158 p->cur_frame_type = SID_FRAME;
159 p->subframe[0].amp_index = get_bits(&gb, 6);
163 /* Extract the info common to both rates */
164 p->cur_rate = info_bits ? RATE_5300 : RATE_6300;
165 p->cur_frame_type = ACTIVE_FRAME;
167 p->pitch_lag[0] = get_bits(&gb, 7);
168 if (p->pitch_lag[0] > 123) /* test if forbidden code */
170 p->pitch_lag[0] += PITCH_MIN;
171 p->subframe[1].ad_cb_lag = get_bits(&gb, 2);
173 p->pitch_lag[1] = get_bits(&gb, 7);
174 if (p->pitch_lag[1] > 123)
176 p->pitch_lag[1] += PITCH_MIN;
177 p->subframe[3].ad_cb_lag = get_bits(&gb, 2);
178 p->subframe[0].ad_cb_lag = 1;
179 p->subframe[2].ad_cb_lag = 1;
181 for (i = 0; i < SUBFRAMES; i++) {
182 /* Extract combined gain */
183 temp = get_bits(&gb, 12);
185 p->subframe[i].dirac_train = 0;
186 if (p->cur_rate == RATE_6300 && p->pitch_lag[i >> 1] < SUBFRAME_LEN - 2) {
187 p->subframe[i].dirac_train = temp >> 11;
191 p->subframe[i].ad_cb_gain = FASTDIV(temp, GAIN_LEVELS);
192 if (p->subframe[i].ad_cb_gain < ad_cb_len) {
193 p->subframe[i].amp_index = temp - p->subframe[i].ad_cb_gain *
200 p->subframe[0].grid_index = get_bits(&gb, 1);
201 p->subframe[1].grid_index = get_bits(&gb, 1);
202 p->subframe[2].grid_index = get_bits(&gb, 1);
203 p->subframe[3].grid_index = get_bits(&gb, 1);
205 if (p->cur_rate == RATE_6300) {
206 skip_bits(&gb, 1); /* skip reserved bit */
208 /* Compute pulse_pos index using the 13-bit combined position index */
209 temp = get_bits(&gb, 13);
210 p->subframe[0].pulse_pos = temp / 810;
212 temp -= p->subframe[0].pulse_pos * 810;
213 p->subframe[1].pulse_pos = FASTDIV(temp, 90);
215 temp -= p->subframe[1].pulse_pos * 90;
216 p->subframe[2].pulse_pos = FASTDIV(temp, 9);
217 p->subframe[3].pulse_pos = temp - p->subframe[2].pulse_pos * 9;
219 p->subframe[0].pulse_pos = (p->subframe[0].pulse_pos << 16) +
221 p->subframe[1].pulse_pos = (p->subframe[1].pulse_pos << 14) +
223 p->subframe[2].pulse_pos = (p->subframe[2].pulse_pos << 16) +
225 p->subframe[3].pulse_pos = (p->subframe[3].pulse_pos << 14) +
228 p->subframe[0].pulse_sign = get_bits(&gb, 6);
229 p->subframe[1].pulse_sign = get_bits(&gb, 5);
230 p->subframe[2].pulse_sign = get_bits(&gb, 6);
231 p->subframe[3].pulse_sign = get_bits(&gb, 5);
232 } else { /* 5300 bps */
233 p->subframe[0].pulse_pos = get_bits(&gb, 12);
234 p->subframe[1].pulse_pos = get_bits(&gb, 12);
235 p->subframe[2].pulse_pos = get_bits(&gb, 12);
236 p->subframe[3].pulse_pos = get_bits(&gb, 12);
238 p->subframe[0].pulse_sign = get_bits(&gb, 4);
239 p->subframe[1].pulse_sign = get_bits(&gb, 4);
240 p->subframe[2].pulse_sign = get_bits(&gb, 4);
241 p->subframe[3].pulse_sign = get_bits(&gb, 4);
248 * Bitexact implementation of sqrt(val/2).
250 static int16_t square_root(int val)
253 int16_t exp = 0x4000;
256 for (i = 0; i < 14; i ++) {
257 int res_exp = res + exp;
258 if (val >= res_exp * res_exp << 1)
266 * Calculate the number of left-shifts required for normalizing the input.
268 * @param num input number
269 * @param width width of the input, 16 bits(0) / 32 bits(1)
271 static int normalize_bits(int num, int width)
273 return width - av_log2(num) - 1;
277 * Scale vector contents based on the largest of their absolutes.
279 static int scale_vector(int16_t *dst, const int16_t *vector, int length)
285 for (i = 0; i < length; i++)
286 max |= FFABS(vector[i]);
288 max = FFMIN(max, 0x7FFF);
289 bits = normalize_bits(max, 15);
291 for (i = 0; i < length; i++)
292 dst[i] = vector[i] << bits >> 3;
298 * Perform inverse quantization of LSP frequencies.
300 * @param cur_lsp the current LSP vector
301 * @param prev_lsp the previous LSP vector
302 * @param lsp_index VQ indices
303 * @param bad_frame bad frame flag
305 static void inverse_quant(int16_t *cur_lsp, int16_t *prev_lsp,
306 uint8_t *lsp_index, int bad_frame)
309 int i, j, temp, stable;
311 /* Check for frame erasure */
318 lsp_index[0] = lsp_index[1] = lsp_index[2] = 0;
321 /* Get the VQ table entry corresponding to the transmitted index */
322 cur_lsp[0] = lsp_band0[lsp_index[0]][0];
323 cur_lsp[1] = lsp_band0[lsp_index[0]][1];
324 cur_lsp[2] = lsp_band0[lsp_index[0]][2];
325 cur_lsp[3] = lsp_band1[lsp_index[1]][0];
326 cur_lsp[4] = lsp_band1[lsp_index[1]][1];
327 cur_lsp[5] = lsp_band1[lsp_index[1]][2];
328 cur_lsp[6] = lsp_band2[lsp_index[2]][0];
329 cur_lsp[7] = lsp_band2[lsp_index[2]][1];
330 cur_lsp[8] = lsp_band2[lsp_index[2]][2];
331 cur_lsp[9] = lsp_band2[lsp_index[2]][3];
333 /* Add predicted vector & DC component to the previously quantized vector */
334 for (i = 0; i < LPC_ORDER; i++) {
335 temp = ((prev_lsp[i] - dc_lsp[i]) * pred + (1 << 14)) >> 15;
336 cur_lsp[i] += dc_lsp[i] + temp;
339 for (i = 0; i < LPC_ORDER; i++) {
340 cur_lsp[0] = FFMAX(cur_lsp[0], 0x180);
341 cur_lsp[LPC_ORDER - 1] = FFMIN(cur_lsp[LPC_ORDER - 1], 0x7e00);
343 /* Stability check */
344 for (j = 1; j < LPC_ORDER; j++) {
345 temp = min_dist + cur_lsp[j - 1] - cur_lsp[j];
348 cur_lsp[j - 1] -= temp;
353 for (j = 1; j < LPC_ORDER; j++) {
354 temp = cur_lsp[j - 1] + min_dist - cur_lsp[j] - 4;
364 memcpy(cur_lsp, prev_lsp, LPC_ORDER * sizeof(*cur_lsp));
368 * Bitexact implementation of 2ab scaled by 1/2^16.
370 * @param a 32 bit multiplicand
371 * @param b 16 bit multiplier
373 #define MULL2(a, b) \
374 ((((a) >> 16) * (b) << 1) + (((a) & 0xffff) * (b) >> 15))
377 * Convert LSP frequencies to LPC coefficients.
379 * @param lpc buffer for LPC coefficients
381 static void lsp2lpc(int16_t *lpc)
383 int f1[LPC_ORDER / 2 + 1];
384 int f2[LPC_ORDER / 2 + 1];
387 /* Calculate negative cosine */
388 for (j = 0; j < LPC_ORDER; j++) {
389 int index = lpc[j] >> 7;
390 int offset = lpc[j] & 0x7f;
391 int temp1 = cos_tab[index] << 16;
392 int temp2 = (cos_tab[index + 1] - cos_tab[index]) *
393 ((offset << 8) + 0x80) << 1;
395 lpc[j] = -(av_sat_dadd32(1 << 15, temp1 + temp2) >> 16);
399 * Compute sum and difference polynomial coefficients
400 * (bitexact alternative to lsp2poly() in lsp.c)
402 /* Initialize with values in Q28 */
404 f1[1] = (lpc[0] << 14) + (lpc[2] << 14);
405 f1[2] = lpc[0] * lpc[2] + (2 << 28);
408 f2[1] = (lpc[1] << 14) + (lpc[3] << 14);
409 f2[2] = lpc[1] * lpc[3] + (2 << 28);
412 * Calculate and scale the coefficients by 1/2 in
413 * each iteration for a final scaling factor of Q25
415 for (i = 2; i < LPC_ORDER / 2; i++) {
416 f1[i + 1] = f1[i - 1] + MULL2(f1[i], lpc[2 * i]);
417 f2[i + 1] = f2[i - 1] + MULL2(f2[i], lpc[2 * i + 1]);
419 for (j = i; j >= 2; j--) {
420 f1[j] = MULL2(f1[j - 1], lpc[2 * i]) +
421 (f1[j] >> 1) + (f1[j - 2] >> 1);
422 f2[j] = MULL2(f2[j - 1], lpc[2 * i + 1]) +
423 (f2[j] >> 1) + (f2[j - 2] >> 1);
428 f1[1] = ((lpc[2 * i] << 16 >> i) + f1[1]) >> 1;
429 f2[1] = ((lpc[2 * i + 1] << 16 >> i) + f2[1]) >> 1;
432 /* Convert polynomial coefficients to LPC coefficients */
433 for (i = 0; i < LPC_ORDER / 2; i++) {
434 int64_t ff1 = f1[i + 1] + f1[i];
435 int64_t ff2 = f2[i + 1] - f2[i];
437 lpc[i] = av_clipl_int32(((ff1 + ff2) << 3) + (1 << 15)) >> 16;
438 lpc[LPC_ORDER - i - 1] = av_clipl_int32(((ff1 - ff2) << 3) +
444 * Quantize LSP frequencies by interpolation and convert them to
445 * the corresponding LPC coefficients.
447 * @param lpc buffer for LPC coefficients
448 * @param cur_lsp the current LSP vector
449 * @param prev_lsp the previous LSP vector
451 static void lsp_interpolate(int16_t *lpc, int16_t *cur_lsp, int16_t *prev_lsp)
454 int16_t *lpc_ptr = lpc;
456 /* cur_lsp * 0.25 + prev_lsp * 0.75 */
457 ff_acelp_weighted_vector_sum(lpc, cur_lsp, prev_lsp,
458 4096, 12288, 1 << 13, 14, LPC_ORDER);
459 ff_acelp_weighted_vector_sum(lpc + LPC_ORDER, cur_lsp, prev_lsp,
460 8192, 8192, 1 << 13, 14, LPC_ORDER);
461 ff_acelp_weighted_vector_sum(lpc + 2 * LPC_ORDER, cur_lsp, prev_lsp,
462 12288, 4096, 1 << 13, 14, LPC_ORDER);
463 memcpy(lpc + 3 * LPC_ORDER, cur_lsp, LPC_ORDER * sizeof(*lpc));
465 for (i = 0; i < SUBFRAMES; i++) {
467 lpc_ptr += LPC_ORDER;
472 * Generate a train of dirac functions with period as pitch lag.
474 static void gen_dirac_train(int16_t *buf, int pitch_lag)
476 int16_t vector[SUBFRAME_LEN];
479 memcpy(vector, buf, SUBFRAME_LEN * sizeof(*vector));
480 for (i = pitch_lag; i < SUBFRAME_LEN; i += pitch_lag) {
481 for (j = 0; j < SUBFRAME_LEN - i; j++)
482 buf[i + j] += vector[j];
487 * Generate fixed codebook excitation vector.
489 * @param vector decoded excitation vector
490 * @param subfrm current subframe
491 * @param cur_rate current bitrate
492 * @param pitch_lag closed loop pitch lag
493 * @param index current subframe index
495 static void gen_fcb_excitation(int16_t *vector, G723_1_Subframe *subfrm,
496 enum Rate cur_rate, int pitch_lag, int index)
500 memset(vector, 0, SUBFRAME_LEN * sizeof(*vector));
502 if (cur_rate == RATE_6300) {
503 if (subfrm->pulse_pos >= max_pos[index])
506 /* Decode amplitudes and positions */
507 j = PULSE_MAX - pulses[index];
508 temp = subfrm->pulse_pos;
509 for (i = 0; i < SUBFRAME_LEN / GRID_SIZE; i++) {
510 temp -= combinatorial_table[j][i];
513 temp += combinatorial_table[j++][i];
514 if (subfrm->pulse_sign & (1 << (PULSE_MAX - j))) {
515 vector[subfrm->grid_index + GRID_SIZE * i] =
516 -fixed_cb_gain[subfrm->amp_index];
518 vector[subfrm->grid_index + GRID_SIZE * i] =
519 fixed_cb_gain[subfrm->amp_index];
524 if (subfrm->dirac_train == 1)
525 gen_dirac_train(vector, pitch_lag);
526 } else { /* 5300 bps */
527 int cb_gain = fixed_cb_gain[subfrm->amp_index];
528 int cb_shift = subfrm->grid_index;
529 int cb_sign = subfrm->pulse_sign;
530 int cb_pos = subfrm->pulse_pos;
531 int offset, beta, lag;
533 for (i = 0; i < 8; i += 2) {
534 offset = ((cb_pos & 7) << 3) + cb_shift + i;
535 vector[offset] = (cb_sign & 1) ? cb_gain : -cb_gain;
540 /* Enhance harmonic components */
541 lag = pitch_contrib[subfrm->ad_cb_gain << 1] + pitch_lag +
542 subfrm->ad_cb_lag - 1;
543 beta = pitch_contrib[(subfrm->ad_cb_gain << 1) + 1];
545 if (lag < SUBFRAME_LEN - 2) {
546 for (i = lag; i < SUBFRAME_LEN; i++)
547 vector[i] += beta * vector[i - lag] >> 15;
553 * Get delayed contribution from the previous excitation vector.
555 static void get_residual(int16_t *residual, int16_t *prev_excitation, int lag)
557 int offset = PITCH_MAX - PITCH_ORDER / 2 - lag;
560 residual[0] = prev_excitation[offset];
561 residual[1] = prev_excitation[offset + 1];
564 for (i = 2; i < SUBFRAME_LEN + PITCH_ORDER - 1; i++)
565 residual[i] = prev_excitation[offset + (i - 2) % lag];
568 static int dot_product(const int16_t *a, const int16_t *b, int length)
572 for (i = 0; i < length; i++) {
573 int prod = a[i] * b[i];
574 sum = av_sat_dadd32(sum, prod);
580 * Generate adaptive codebook excitation.
582 static void gen_acb_excitation(int16_t *vector, int16_t *prev_excitation,
583 int pitch_lag, G723_1_Subframe *subfrm,
586 int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
587 const int16_t *cb_ptr;
588 int lag = pitch_lag + subfrm->ad_cb_lag - 1;
593 get_residual(residual, prev_excitation, lag);
595 /* Select quantization table */
596 if (cur_rate == RATE_6300 && pitch_lag < SUBFRAME_LEN - 2)
597 cb_ptr = adaptive_cb_gain85;
599 cb_ptr = adaptive_cb_gain170;
601 /* Calculate adaptive vector */
602 cb_ptr += subfrm->ad_cb_gain * 20;
603 for (i = 0; i < SUBFRAME_LEN; i++) {
604 sum = dot_product(residual + i, cb_ptr, PITCH_ORDER);
605 vector[i] = av_sat_dadd32(1 << 15, sum) >> 16;
610 * Estimate maximum auto-correlation around pitch lag.
612 * @param buf buffer with offset applied
613 * @param offset offset of the excitation vector
614 * @param ccr_max pointer to the maximum auto-correlation
615 * @param pitch_lag decoded pitch lag
616 * @param length length of autocorrelation
617 * @param dir forward lag(1) / backward lag(-1)
619 static int autocorr_max(const int16_t *buf, int offset, int *ccr_max,
620 int pitch_lag, int length, int dir)
622 int limit, ccr, lag = 0;
625 pitch_lag = FFMIN(PITCH_MAX - 3, pitch_lag);
627 limit = FFMIN(FRAME_LEN + PITCH_MAX - offset - length, pitch_lag + 3);
629 limit = pitch_lag + 3;
631 for (i = pitch_lag - 3; i <= limit; i++) {
632 ccr = dot_product(buf, buf + dir * i, length);
634 if (ccr > *ccr_max) {
643 * Calculate pitch postfilter optimal and scaling gains.
645 * @param lag pitch postfilter forward/backward lag
646 * @param ppf pitch postfilter parameters
647 * @param cur_rate current bitrate
648 * @param tgt_eng target energy
649 * @param ccr cross-correlation
650 * @param res_eng residual energy
652 static void comp_ppf_gains(int lag, PPFParam *ppf, enum Rate cur_rate,
653 int tgt_eng, int ccr, int res_eng)
655 int pf_residual; /* square of postfiltered residual */
660 temp1 = tgt_eng * res_eng >> 1;
661 temp2 = ccr * ccr << 1;
664 if (ccr >= res_eng) {
665 ppf->opt_gain = ppf_gain_weight[cur_rate];
667 ppf->opt_gain = (ccr << 15) / res_eng *
668 ppf_gain_weight[cur_rate] >> 15;
670 /* pf_res^2 = tgt_eng + 2*ccr*gain + res_eng*gain^2 */
671 temp1 = (tgt_eng << 15) + (ccr * ppf->opt_gain << 1);
672 temp2 = (ppf->opt_gain * ppf->opt_gain >> 15) * res_eng;
673 pf_residual = av_sat_add32(temp1, temp2 + (1 << 15)) >> 16;
675 if (tgt_eng >= pf_residual << 1) {
678 temp1 = (tgt_eng << 14) / pf_residual;
681 /* scaling_gain = sqrt(tgt_eng/pf_res^2) */
682 ppf->sc_gain = square_root(temp1 << 16);
685 ppf->sc_gain = 0x7fff;
688 ppf->opt_gain = av_clip_int16(ppf->opt_gain * ppf->sc_gain >> 15);
692 * Calculate pitch postfilter parameters.
694 * @param p the context
695 * @param offset offset of the excitation vector
696 * @param pitch_lag decoded pitch lag
697 * @param ppf pitch postfilter parameters
698 * @param cur_rate current bitrate
700 static void comp_ppf_coeff(G723_1_Context *p, int offset, int pitch_lag,
701 PPFParam *ppf, enum Rate cur_rate)
710 * 1 - forward cross-correlation
711 * 2 - forward residual energy
712 * 3 - backward cross-correlation
713 * 4 - backward residual energy
715 int energy[5] = {0, 0, 0, 0, 0};
716 int16_t *buf = p->audio + LPC_ORDER + offset;
717 int fwd_lag = autocorr_max(buf, offset, &energy[1], pitch_lag,
719 int back_lag = autocorr_max(buf, offset, &energy[3], pitch_lag,
724 ppf->sc_gain = 0x7fff;
726 /* Case 0, Section 3.6 */
727 if (!back_lag && !fwd_lag)
730 /* Compute target energy */
731 energy[0] = dot_product(buf, buf, SUBFRAME_LEN);
733 /* Compute forward residual energy */
735 energy[2] = dot_product(buf + fwd_lag, buf + fwd_lag, SUBFRAME_LEN);
737 /* Compute backward residual energy */
739 energy[4] = dot_product(buf - back_lag, buf - back_lag, SUBFRAME_LEN);
741 /* Normalize and shorten */
743 for (i = 0; i < 5; i++)
744 temp1 = FFMAX(energy[i], temp1);
746 scale = normalize_bits(temp1, 31);
747 for (i = 0; i < 5; i++)
748 energy[i] = (energy[i] << scale) >> 16;
750 if (fwd_lag && !back_lag) { /* Case 1 */
751 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
753 } else if (!fwd_lag) { /* Case 2 */
754 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
756 } else { /* Case 3 */
759 * Select the largest of energy[1]^2/energy[2]
760 * and energy[3]^2/energy[4]
762 temp1 = energy[4] * ((energy[1] * energy[1] + (1 << 14)) >> 15);
763 temp2 = energy[2] * ((energy[3] * energy[3] + (1 << 14)) >> 15);
764 if (temp1 >= temp2) {
765 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
768 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
775 * Classify frames as voiced/unvoiced.
777 * @param p the context
778 * @param pitch_lag decoded pitch_lag
779 * @param exc_eng excitation energy estimation
780 * @param scale scaling factor of exc_eng
782 * @return residual interpolation index if voiced, 0 otherwise
784 static int comp_interp_index(G723_1_Context *p, int pitch_lag,
785 int *exc_eng, int *scale)
787 int offset = PITCH_MAX + 2 * SUBFRAME_LEN;
788 int16_t *buf = p->audio + LPC_ORDER;
790 int index, ccr, tgt_eng, best_eng, temp;
792 *scale = scale_vector(buf, p->excitation, FRAME_LEN + PITCH_MAX);
795 /* Compute maximum backward cross-correlation */
797 index = autocorr_max(buf, offset, &ccr, pitch_lag, SUBFRAME_LEN * 2, -1);
798 ccr = av_sat_add32(ccr, 1 << 15) >> 16;
800 /* Compute target energy */
801 tgt_eng = dot_product(buf, buf, SUBFRAME_LEN * 2);
802 *exc_eng = av_sat_add32(tgt_eng, 1 << 15) >> 16;
807 /* Compute best energy */
808 best_eng = dot_product(buf - index, buf - index, SUBFRAME_LEN * 2);
809 best_eng = av_sat_add32(best_eng, 1 << 15) >> 16;
811 temp = best_eng * *exc_eng >> 3;
813 if (temp < ccr * ccr)
820 * Peform residual interpolation based on frame classification.
822 * @param buf decoded excitation vector
823 * @param out output vector
824 * @param lag decoded pitch lag
825 * @param gain interpolated gain
826 * @param rseed seed for random number generator
828 static void residual_interp(int16_t *buf, int16_t *out, int lag,
829 int gain, int *rseed)
832 if (lag) { /* Voiced */
833 int16_t *vector_ptr = buf + PITCH_MAX;
835 for (i = 0; i < lag; i++)
836 out[i] = vector_ptr[i - lag] * 3 >> 2;
837 av_memcpy_backptr((uint8_t*)(out + lag), lag * sizeof(*out),
838 (FRAME_LEN - lag) * sizeof(*out));
839 } else { /* Unvoiced */
840 for (i = 0; i < FRAME_LEN; i++) {
841 *rseed = *rseed * 521 + 259;
842 out[i] = gain * *rseed >> 15;
844 memset(buf, 0, (FRAME_LEN + PITCH_MAX) * sizeof(*buf));
849 * Perform IIR filtering.
851 * @param fir_coef FIR coefficients
852 * @param iir_coef IIR coefficients
853 * @param src source vector
854 * @param dest destination vector
856 static inline void iir_filter(int16_t *fir_coef, int16_t *iir_coef,
857 int16_t *src, int *dest)
861 for (m = 0; m < SUBFRAME_LEN; m++) {
863 for (n = 1; n <= LPC_ORDER; n++) {
864 filter -= fir_coef[n - 1] * src[m - n] -
865 iir_coef[n - 1] * (dest[m - n] >> 16);
868 dest[m] = av_clipl_int32((src[m] << 16) + (filter << 3) + (1 << 15));
873 * Adjust gain of postfiltered signal.
875 * @param p the context
876 * @param buf postfiltered output vector
877 * @param energy input energy coefficient
879 static void gain_scale(G723_1_Context *p, int16_t * buf, int energy)
881 int num, denom, gain, bits1, bits2;
886 for (i = 0; i < SUBFRAME_LEN; i++) {
887 int temp = buf[i] >> 2;
889 denom = av_sat_dadd32(denom, temp);
893 bits1 = normalize_bits(num, 31);
894 bits2 = normalize_bits(denom, 31);
895 num = num << bits1 >> 1;
898 bits2 = 5 + bits1 - bits2;
899 bits2 = FFMAX(0, bits2);
901 gain = (num >> 1) / (denom >> 16);
902 gain = square_root(gain << 16 >> bits2);
907 for (i = 0; i < SUBFRAME_LEN; i++) {
908 p->pf_gain = (15 * p->pf_gain + gain + (1 << 3)) >> 4;
909 buf[i] = av_clip_int16((buf[i] * (p->pf_gain + (p->pf_gain >> 4)) +
915 * Perform formant filtering.
917 * @param p the context
918 * @param lpc quantized lpc coefficients
919 * @param buf input buffer
920 * @param dst output buffer
922 static void formant_postfilter(G723_1_Context *p, int16_t *lpc,
923 int16_t *buf, int16_t *dst)
925 int16_t filter_coef[2][LPC_ORDER];
926 int filter_signal[LPC_ORDER + FRAME_LEN], *signal_ptr;
929 memcpy(buf, p->fir_mem, LPC_ORDER * sizeof(*buf));
930 memcpy(filter_signal, p->iir_mem, LPC_ORDER * sizeof(*filter_signal));
932 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
933 for (k = 0; k < LPC_ORDER; k++) {
934 filter_coef[0][k] = (-lpc[k] * postfilter_tbl[0][k] +
936 filter_coef[1][k] = (-lpc[k] * postfilter_tbl[1][k] +
939 iir_filter(filter_coef[0], filter_coef[1], buf + i,
944 memcpy(p->fir_mem, buf + FRAME_LEN, LPC_ORDER * sizeof(*p->fir_mem));
945 memcpy(p->iir_mem, filter_signal + FRAME_LEN,
946 LPC_ORDER * sizeof(*p->iir_mem));
949 signal_ptr = filter_signal + LPC_ORDER;
950 for (i = 0; i < SUBFRAMES; i++) {
956 scale = scale_vector(dst, buf, SUBFRAME_LEN);
958 /* Compute auto correlation coefficients */
959 auto_corr[0] = dot_product(dst, dst + 1, SUBFRAME_LEN - 1);
960 auto_corr[1] = dot_product(dst, dst, SUBFRAME_LEN);
962 /* Compute reflection coefficient */
963 temp = auto_corr[1] >> 16;
965 temp = (auto_corr[0] >> 2) / temp;
967 p->reflection_coef = (3 * p->reflection_coef + temp + 2) >> 2;
968 temp = -p->reflection_coef >> 1 & ~3;
970 /* Compensation filter */
971 for (j = 0; j < SUBFRAME_LEN; j++) {
972 dst[j] = av_sat_dadd32(signal_ptr[j],
973 (signal_ptr[j - 1] >> 16) * temp) >> 16;
976 /* Compute normalized signal energy */
977 temp = 2 * scale + 4;
979 energy = av_clipl_int32((int64_t)auto_corr[1] << -temp);
981 energy = auto_corr[1] >> temp;
983 gain_scale(p, dst, energy);
986 signal_ptr += SUBFRAME_LEN;
991 static int sid_gain_to_lsp_index(int gain)
995 else if (gain < 0x20)
996 return gain - 8 << 7;
998 return gain - 20 << 8;
1001 static inline int cng_rand(int *state, int base)
1003 *state = (*state * 521 + 259) & 0xFFFF;
1004 return (*state & 0x7FFF) * base >> 15;
1007 static int estimate_sid_gain(G723_1_Context *p)
1009 int i, shift, seg, seg2, t, val, val_add, x, y;
1011 shift = 16 - p->cur_gain * 2;
1013 t = p->sid_gain << shift;
1015 t = p->sid_gain >> -shift;
1016 x = t * cng_filt[0] >> 16;
1018 if (x >= cng_bseg[2])
1021 if (x >= cng_bseg[1]) {
1026 seg = (x >= cng_bseg[0]);
1028 seg2 = FFMIN(seg, 3);
1032 for (i = 0; i < shift; i++) {
1033 t = seg * 32 + (val << seg2);
1042 t = seg * 32 + (val << seg2);
1045 t = seg * 32 + (val + 1 << seg2);
1047 val = (seg2 - 1 << 4) + val;
1051 t = seg * 32 + (val - 1 << seg2);
1053 val = (seg2 - 1 << 4) + val;
1061 static void generate_noise(G723_1_Context *p)
1065 int signs[SUBFRAMES / 2 * 11], pos[SUBFRAMES / 2 * 11];
1066 int tmp[SUBFRAME_LEN * 2];
1067 int16_t *vector_ptr;
1069 int b0, c, delta, x, shift;
1071 p->pitch_lag[0] = cng_rand(&p->cng_random_seed, 21) + 123;
1072 p->pitch_lag[1] = cng_rand(&p->cng_random_seed, 19) + 123;
1074 for (i = 0; i < SUBFRAMES; i++) {
1075 p->subframe[i].ad_cb_gain = cng_rand(&p->cng_random_seed, 50) + 1;
1076 p->subframe[i].ad_cb_lag = cng_adaptive_cb_lag[i];
1079 for (i = 0; i < SUBFRAMES / 2; i++) {
1080 t = cng_rand(&p->cng_random_seed, 1 << 13);
1082 off[i * 2 + 1] = ((t >> 1) & 1) + SUBFRAME_LEN;
1084 for (j = 0; j < 11; j++) {
1085 signs[i * 11 + j] = (t & 1) * 2 - 1 << 14;
1091 for (i = 0; i < SUBFRAMES; i++) {
1092 for (j = 0; j < SUBFRAME_LEN / 2; j++)
1094 t = SUBFRAME_LEN / 2;
1095 for (j = 0; j < pulses[i]; j++, idx++) {
1096 int idx2 = cng_rand(&p->cng_random_seed, t);
1098 pos[idx] = tmp[idx2] * 2 + off[i];
1099 tmp[idx2] = tmp[--t];
1103 vector_ptr = p->audio + LPC_ORDER;
1104 memcpy(vector_ptr, p->prev_excitation,
1105 PITCH_MAX * sizeof(*p->excitation));
1106 for (i = 0; i < SUBFRAMES; i += 2) {
1107 gen_acb_excitation(vector_ptr, vector_ptr,
1108 p->pitch_lag[i >> 1], &p->subframe[i],
1110 gen_acb_excitation(vector_ptr + SUBFRAME_LEN,
1111 vector_ptr + SUBFRAME_LEN,
1112 p->pitch_lag[i >> 1], &p->subframe[i + 1],
1116 for (j = 0; j < SUBFRAME_LEN * 2; j++)
1117 t |= FFABS(vector_ptr[j]);
1118 t = FFMIN(t, 0x7FFF);
1122 shift = -10 + av_log2(t);
1128 for (j = 0; j < SUBFRAME_LEN * 2; j++) {
1129 t = vector_ptr[j] << -shift;
1134 for (j = 0; j < SUBFRAME_LEN * 2; j++) {
1135 t = vector_ptr[j] >> shift;
1142 for (j = 0; j < 11; j++)
1143 b0 += tmp[pos[(i / 2) * 11 + j]] * signs[(i / 2) * 11 + j];
1144 b0 = b0 * 2 * 2979LL + (1 << 29) >> 30; // approximated division by 11
1146 c = p->cur_gain * (p->cur_gain * SUBFRAME_LEN >> 5);
1147 if (shift * 2 + 3 >= 0)
1148 c >>= shift * 2 + 3;
1150 c <<= -(shift * 2 + 3);
1151 c = (av_clipl_int32(sum << 1) - c) * 2979LL >> 15;
1153 delta = b0 * b0 * 2 - c;
1157 delta = square_root(delta);
1160 if (FFABS(t) < FFABS(x))
1168 x = av_clip(x, -10000, 10000);
1170 for (j = 0; j < 11; j++) {
1171 idx = (i / 2) * 11 + j;
1172 vector_ptr[pos[idx]] = av_clip_int16(vector_ptr[pos[idx]] +
1173 (x * signs[idx] >> 15));
1176 /* copy decoded data to serve as a history for the next decoded subframes */
1177 memcpy(vector_ptr + PITCH_MAX, vector_ptr,
1178 sizeof(*vector_ptr) * SUBFRAME_LEN * 2);
1179 vector_ptr += SUBFRAME_LEN * 2;
1181 /* Save the excitation for the next frame */
1182 memcpy(p->prev_excitation, p->audio + LPC_ORDER + FRAME_LEN,
1183 PITCH_MAX * sizeof(*p->excitation));
1186 static int g723_1_decode_frame(AVCodecContext *avctx, void *data,
1187 int *got_frame_ptr, AVPacket *avpkt)
1189 G723_1_Context *p = avctx->priv_data;
1190 AVFrame *frame = data;
1191 const uint8_t *buf = avpkt->data;
1192 int buf_size = avpkt->size;
1193 int dec_mode = buf[0] & 3;
1195 PPFParam ppf[SUBFRAMES];
1196 int16_t cur_lsp[LPC_ORDER];
1197 int16_t lpc[SUBFRAMES * LPC_ORDER];
1198 int16_t acb_vector[SUBFRAME_LEN];
1200 int bad_frame = 0, i, j, ret;
1201 int16_t *audio = p->audio;
1203 if (buf_size < frame_size[dec_mode]) {
1205 av_log(avctx, AV_LOG_WARNING,
1206 "Expected %d bytes, got %d - skipping packet\n",
1207 frame_size[dec_mode], buf_size);
1212 if (unpack_bitstream(p, buf, buf_size) < 0) {
1214 if (p->past_frame_type == ACTIVE_FRAME)
1215 p->cur_frame_type = ACTIVE_FRAME;
1217 p->cur_frame_type = UNTRANSMITTED_FRAME;
1220 frame->nb_samples = FRAME_LEN;
1221 if ((ret = ff_get_buffer(avctx, frame, 0)) < 0) {
1222 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1226 out = (int16_t *)frame->data[0];
1228 if (p->cur_frame_type == ACTIVE_FRAME) {
1230 p->erased_frames = 0;
1231 else if (p->erased_frames != 3)
1234 inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, bad_frame);
1235 lsp_interpolate(lpc, cur_lsp, p->prev_lsp);
1237 /* Save the lsp_vector for the next frame */
1238 memcpy(p->prev_lsp, cur_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
1240 /* Generate the excitation for the frame */
1241 memcpy(p->excitation, p->prev_excitation,
1242 PITCH_MAX * sizeof(*p->excitation));
1243 if (!p->erased_frames) {
1244 int16_t *vector_ptr = p->excitation + PITCH_MAX;
1246 /* Update interpolation gain memory */
1247 p->interp_gain = fixed_cb_gain[(p->subframe[2].amp_index +
1248 p->subframe[3].amp_index) >> 1];
1249 for (i = 0; i < SUBFRAMES; i++) {
1250 gen_fcb_excitation(vector_ptr, &p->subframe[i], p->cur_rate,
1251 p->pitch_lag[i >> 1], i);
1252 gen_acb_excitation(acb_vector, &p->excitation[SUBFRAME_LEN * i],
1253 p->pitch_lag[i >> 1], &p->subframe[i],
1255 /* Get the total excitation */
1256 for (j = 0; j < SUBFRAME_LEN; j++) {
1257 int v = av_clip_int16(vector_ptr[j] << 1);
1258 vector_ptr[j] = av_clip_int16(v + acb_vector[j]);
1260 vector_ptr += SUBFRAME_LEN;
1263 vector_ptr = p->excitation + PITCH_MAX;
1265 p->interp_index = comp_interp_index(p, p->pitch_lag[1],
1266 &p->sid_gain, &p->cur_gain);
1268 /* Peform pitch postfiltering */
1269 if (p->postfilter) {
1271 for (j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1272 comp_ppf_coeff(p, i, p->pitch_lag[j >> 1],
1273 ppf + j, p->cur_rate);
1275 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1276 ff_acelp_weighted_vector_sum(p->audio + LPC_ORDER + i,
1278 vector_ptr + i + ppf[j].index,
1281 1 << 14, 15, SUBFRAME_LEN);
1283 audio = vector_ptr - LPC_ORDER;
1286 /* Save the excitation for the next frame */
1287 memcpy(p->prev_excitation, p->excitation + FRAME_LEN,
1288 PITCH_MAX * sizeof(*p->excitation));
1290 p->interp_gain = (p->interp_gain * 3 + 2) >> 2;
1291 if (p->erased_frames == 3) {
1293 memset(p->excitation, 0,
1294 (FRAME_LEN + PITCH_MAX) * sizeof(*p->excitation));
1295 memset(p->prev_excitation, 0,
1296 PITCH_MAX * sizeof(*p->excitation));
1297 memset(frame->data[0], 0,
1298 (FRAME_LEN + LPC_ORDER) * sizeof(int16_t));
1300 int16_t *buf = p->audio + LPC_ORDER;
1302 /* Regenerate frame */
1303 residual_interp(p->excitation, buf, p->interp_index,
1304 p->interp_gain, &p->random_seed);
1306 /* Save the excitation for the next frame */
1307 memcpy(p->prev_excitation, buf + (FRAME_LEN - PITCH_MAX),
1308 PITCH_MAX * sizeof(*p->excitation));
1311 p->cng_random_seed = CNG_RANDOM_SEED;
1313 if (p->cur_frame_type == SID_FRAME) {
1314 p->sid_gain = sid_gain_to_lsp_index(p->subframe[0].amp_index);
1315 inverse_quant(p->sid_lsp, p->prev_lsp, p->lsp_index, 0);
1316 } else if (p->past_frame_type == ACTIVE_FRAME) {
1317 p->sid_gain = estimate_sid_gain(p);
1320 if (p->past_frame_type == ACTIVE_FRAME)
1321 p->cur_gain = p->sid_gain;
1323 p->cur_gain = (p->cur_gain * 7 + p->sid_gain) >> 3;
1325 lsp_interpolate(lpc, p->sid_lsp, p->prev_lsp);
1326 /* Save the lsp_vector for the next frame */
1327 memcpy(p->prev_lsp, p->sid_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
1330 p->past_frame_type = p->cur_frame_type;
1332 memcpy(p->audio, p->synth_mem, LPC_ORDER * sizeof(*p->audio));
1333 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1334 ff_celp_lp_synthesis_filter(p->audio + i, &lpc[j * LPC_ORDER],
1335 audio + i, SUBFRAME_LEN, LPC_ORDER,
1337 memcpy(p->synth_mem, p->audio + FRAME_LEN, LPC_ORDER * sizeof(*p->audio));
1339 if (p->postfilter) {
1340 formant_postfilter(p, lpc, p->audio, out);
1341 } else { // if output is not postfiltered it should be scaled by 2
1342 for (i = 0; i < FRAME_LEN; i++)
1343 out[i] = av_clip_int16(p->audio[LPC_ORDER + i] << 1);
1348 return frame_size[dec_mode];
1351 #define OFFSET(x) offsetof(G723_1_Context, x)
1352 #define AD AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1354 static const AVOption options[] = {
1355 { "postfilter", "postfilter on/off", OFFSET(postfilter), AV_OPT_TYPE_INT,
1356 { .i64 = 1 }, 0, 1, AD },
1361 static const AVClass g723_1dec_class = {
1362 .class_name = "G.723.1 decoder",
1363 .item_name = av_default_item_name,
1365 .version = LIBAVUTIL_VERSION_INT,
1368 AVCodec ff_g723_1_decoder = {
1370 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
1371 .type = AVMEDIA_TYPE_AUDIO,
1372 .id = AV_CODEC_ID_G723_1,
1373 .priv_data_size = sizeof(G723_1_Context),
1374 .init = g723_1_decode_init,
1375 .decode = g723_1_decode_frame,
1376 .capabilities = CODEC_CAP_SUBFRAMES | CODEC_CAP_DR1,
1377 .priv_class = &g723_1dec_class,